Relationship between cortical state and spiking activity in the lateral geniculate nucleus of marmosets

Abstract

Key points

• How parallel are the primate visual pathways? In the present study, we demonstrate that parallel visual pathways in the dorsal lateral geniculate nucleus (LGN) show distinct patterns of interaction with rhythmic activity in the primary visual cortex (V1).

• In the V1 of anaesthetized marmosets, the EEG frequency spectrum undergoes transient changes that are characterized by fluctuations in delta‐band EEG power.

• We show that, on multisecond timescales, spiking activity in an evolutionary primitive (koniocellular) LGN pathway is specifically linked to these slow EEG spectrum changes. By contrast, on subsecond (delta frequency) timescales, cortical oscillations can entrain spiking activity throughout the entire LGN.

• Our results are consistent with the hypothesis that, in waking animals, the koniocellular pathway selectively participates in brain circuits controlling vigilance and attention.

Abstract

The major afferent cortical pathway in the visual system passes through the dorsal lateral geniculate nucleus (LGN), where nerve signals originating in the eye can first interact with brain circuits regulating visual processing, vigilance and attention. In the present study, we investigated how ongoing and visually driven activity in magnocellular (M), parvocellular (P) and koniocellular (K) layers of the LGN are related to cortical state. We recorded extracellular spiking activity in the LGN simultaneously with local field potentials (LFP) in primary visual cortex, in sufentanil‐anaesthetized marmoset monkeys. We found that asynchronous cortical states (marked by low power in delta‐band LFPs) are linked to high spike rates in K cells (but not P cells or M cells), on multisecond timescales. Cortical asynchrony precedes the increases in K cell spike rates by 1–3 s, implying causality. At subsecond timescales, the spiking activity in many cells of all (M, P and K) classes is phase‐locked to delta waves in the cortical LFP, and more cells are phase‐locked during synchronous cortical states than during asynchronous cortical states. The switch from low‐to‐high spike rates in K cells does not degrade their visual signalling capacity. By contrast, during asynchronous cortical states, the fidelity of visual signals transmitted by K cells is improved, probably because K cell responses become less rectified. Overall, the data show that slow fluctuations in cortical state are selectively linked to K pathway spiking activity, whereas delta‐frequency cortical oscillations entrain spiking activity throughout the entire LGN, in anaesthetized marmosets.

Key points

• How parallel are the primate visual pathways? In the present study, we demonstrate that parallel visual pathways in the dorsal lateral geniculate nucleus (LGN) show distinct patterns of interaction with rhythmic activity in the primary visual cortex (V1).

• In the V1 of anaesthetized marmosets, the EEG frequency spectrum undergoes transient changes that are characterized by fluctuations in delta‐band EEG power.

• We show that, on multisecond timescales, spiking activity in an evolutionary primitive (koniocellular) LGN pathway is specifically linked to these slow EEG spectrum changes. By contrast, on subsecond (delta frequency) timescales, cortical oscillations can entrain spiking activity throughout the entire LGN.

• Our results are consistent with the hypothesis that, in waking animals, the koniocellular pathway selectively participates in brain circuits controlling vigilance and attention.

Abbreviations

AUC

area under the curve

CRT

cathode ray tube

K

koniocellular

K‐bon

koniocellular blue‐on/yellow off

LFP

local field potential

LGN

lateral geniculate nucleus

M

magnocellular

ML

medium/long‐wave sensitive

P

parvocellular

ROC

receiver operator characteristic

S

short‐wave sensitive

V1

primary visual cortex

Introduction

The magnocellular (M), parvocellular (P) and koniocellular (K) pathways form three cortical afferent visual streams in primates (Casagrande, 1994; Hendry and Reid, 2000; Nassi & Callaway, 2009). Visual signals are considered to travel in parallel on these three streams through the lateral geniculate nucleus (LGN) en route to the visual cortex (V1), where cortical circuits can use these inputs to extract complex features from the visual image.

In addition to receiving this feed‐forward visual signal flow, both the LGN and V1 are connected to brainstem centres regulating vigilance and sleep–wake cycles (Bickford et al. 2000; Steriade, 2003; Sherman & Guillery, 2006; Jones, 2007). Furthermore, there are substantial reciprocal projections from layer 6 of V1 to LGN. These projections can exert a direct and indirect (via the thalamic reticular nucleus) influence on LGN cells (Sherman & Guillery, 2006; Sherman, 2016). Finally, and most importantly, the intrinsic circuitry of V1 generates oscillatory states ranging from synchronous (coherent low‐frequency activity) to asynchronous (less coherent) (Buzsáki & Draguhn, 2004; Wang, 2010). Unsurprisingly, studies in rodents, primates and carnivores therefore agree that thalamocortical pathways exhibit brain state‐dependent changes in activity, across a range of EEG band frequencies (McCormick et al. 1992; Contreras & Steriade, 1995; Destexhe et al. 1998; Crunelli & Hughes, 2010; Zagha & McCormick, 2014).

Briggs & Usrey (2009) showed that direct reciprocal pathways from cortex to LGN are organized in parallel with properties analogous to the M, P and K input streams. This result implies that thalamocortical loops comprise independent parallel subsystems, as illustrated schematically in Fig. 1 A. However, a different possibility is suggested by studies in comparative anatomy. The K stream is part of an evolutionary ancient group of thalamocortical pathways including, for example, paralemniscal somatosensory and tegmental auditory pathways (Casagrande, 1994; Jones, 2001, 2007). The cortical projections of K cells are more diverse and widespread than those of P and M cells, and projection targets of K cells include supragranular layers of primary visual cortex, as well as extrastriate cortices (Fitzpatrick et al. 1983; Casagrande, 1994; Sincich et al. 2004; Casagrande et al. 2007; Jones, 2007; Warner et al. 2010). These anatomical features imply that cortical activity generated by K pathways could feed into activity generated by P and M pathways, as shown schematically in Fig. 1 B.

Figure 1. Illustration showing alternative scenarios for organization of reciprocal circuits between the LGN and V1.

Many connections (e.g. with the thalamic reticular nucleus) are omitted for clarity. A, parallel reciprocal pathways connect P, K and M layers of the LGN to distinct non‐interacting circuits in V1. B, parallel afferent pathways engage with linked cortical circuits enabling functional interaction between P, K and M pathways at the cortical level, and in thalamocortical pathways to the LGN. [Color figure can be viewed at wileyonlinelibrary.com]

In sufentanil‐anaesthetized marmosets, the V1 displays infrequent and short‐lived periods of asynchronous (low delta‐power) EEG activity (Cheong et al. 2011). These V1 state changes are analogous to changes seen when awake animals move from inattentive to attentive states. We previously demonstrated that K cells (but not P cells or M cells) show coherent slow sustained increases in spike discharge rate (Cheong et al. 2011). In the present study, we use time‐series analyses and visual stimuli to investigate the temporal relationship between V1 state and coherent activity in LGN. We also investigate whether and how V1 state can influence visual signal transmission through the LGN.

Methods

Ethical approval

Procedures were approved by institutional (University of Melbourne and University of Sydney) Animal Experimentation and Ethics Committees. Procedures conform with the Society for Neuroscience and Australian National Health and Medical Research Council policies on the use of animals in neuroscience research, as well as with the reporting guidelines of The Journal of Physiology (Grundy, 2015).

Animal preparation

Extracellular recordings were made from the LGN of 13 adult marmosets (Callithrix jacchus, weight range 335–420 g) obtained from the Australian National Non‐Human Primate Breeding Facility in Gippsland, Victoria, Australia. The visual‐evoked response properties of many cells from these animals have been reported previously (Cheong et al. 2013; Cheong and Pietersen, 2014; Pietersen et al. 2014); all responses were re‐analysed for the present study. Anaesthesia was induced by i.m. injection of 12 mg kg⁻¹ Alfaxan (Jurox, NSW, Australia) and 3 mg kg⁻¹ diazepam (Roche, NSW, Australia). During surgery, anaesthesia levels were maintained with supplementary doses of Alfaxan. Local anaesthesia (Xylocaine 2%; AstraZeneca, NSW, Australia) was applied to the surgery sites. A tail vein was cannulized and an endotracheal tube was inserted. The animal was placed in a stereotaxic frame and craniotomies were made over the right LGN and primary visual cortex. Animals were artificially respired with a 70%:30% mixture of NO₂:carbogen (5% CO₂ in O₂). Anaesthesia and analgesia were maintained during recording by i.v. sufentanil citrate infusion (6–30 μg kg⁻¹ h⁻¹; Sufenta Forte; Janssen Cilag, Beerse, Belgium) in physiological solution (sodium lactate; Baxter International, NSW, Australia) with added dexamethasone (0.4 mg kg⁻¹ h⁻¹; Mayne Pharma, VIC, Australia) and Synthamin 17 (amino acids 10%; 225 mg kg⁻¹ h⁻¹; Baxter International). Following establishment of a stable anaesthetic plane, pancuronium bromide (0.3 mg kg⁻¹ h⁻¹; AstraZeneca) was added to the infusion solution to induce and maintain muscular paralysis. The respiratory depressant effect of sufentanil obviated the need for a loading dose of paralytic. The animal was artificially ventilated to maintain the end‐tidal CO₂ close to 3.7%. EEG and ECG signals were monitored to ensure adequate anaesthesia levels. Dominance of low frequencies (<10 Hz) in the EEG recording and stability of the EEG frequency spectrum under intermittent noxious stimulus (toe‐pinch) were taken as an indication of adequate anaesthesia. We found that low sufentanil dose rates in the range cited above were always very effective during the first 24 h of recordings. Thereafter, drifts towards higher frequencies (>10 Hz) in the EEG record were counteracted by increasing the rate of venous infusion or the concentration of anaesthetic. The typical duration of a recording session was 72–94 h. Rectal temperature was maintained near 37.0 °C with a thermistor‐controlled heating blanket. An hourly log of the vital signs described above was maintained. Daily additional antibiotics (80–120 mg kg^–1 procaine penicillin; Norocillin SA, Norbrook, UK) were injected i.m. Pupils were dilated with phenylephrine hydrochloride (10%; Chauvin Pharmaceuticals, Kingston‐upon‐Thames, UK). Corneas were protected with gas‐permeable contact lenses that normally remained in place for the duration of the experiment. Supplementary lenses (with power determined by maximizing the spatial resolution of the first receptive fields encountered for each eye) were used to focus the eyes at a distance of 114 cm. As reported previously (Buzás et al. 2006), we saw little evidence for drifts in eye position or refractive state over the duration of the experiment.

Recording procedures

A recording electrode (parylene‐insulated stainless steel; 5–11 MΩ; FHC Inc., Bowdoin, MN, USA) was lowered into the LGN through a guide tube using a hydraulic micropositioner (model 640; David Kopf Instruments, Tujunga, CA, USA). Action potential waveforms of single cells were discriminated by principal component analysis of amplified voltage signals. Extracellular spike events were captured at 100 μs resolution using a Power Mac G5 (Apple, Cupertino, CA, USA) running open‐source data acquisition software (EXPO/OpenGL; P. Lennie, Rochester, NY, USA).

Bipolar electrodes, constructed from 0.05 mm diameter lacquer‐insulated nickel‐chromium wire, were inserted into V1 to record local field potentials (LFP). Each electrode had a vertical tip separation of ∼0.5 mm. The bandpass filtered LFP (0.3–300 Hz) signal was amplified and displayed on a digital storage oscilloscope (DPO2014; Tektronix, Beaverton, OR, USA). A large uniform achromatic flashing stimulus was presented and electrodes were advanced until clear stimulus‐locked modulation of the LFP was evident. Histological reconstruction (see below) showed the lower tip of the electrode was typically ∼1.0 mm below the cortical surface. The LFP was displayed and saved simultaneously with the raw spike signal recorded in the LGN on the oscilloscope in epochs of 40 s at sample rate 3.125 kHz.

Visual stimuli

A front‐silvered gimballed mirror was used to bring the receptive field onto the centre of a cathode ray tube (CRT) monitor (G520; Sony, Tokyo, Japan; 100 Hz refresh rate). For each phosphor, the relationship between the output of the video card and the photopic luminance was determined. The inverse of this relationship was applied to the signals that were sent to the video card. Visual stimuli were generated using the same computer and software that collected spike waveforms. Stimuli were presented on a grey background (guns set to half‐maximum intensity) at a mean luminance close to 50 cd m⁻² and mean chromaticity x = 0.361, y = 0.363. Visual stimuli were presented through the dominant eye only. Stimuli were placed as close as practicable to the centre of the screen aiming to reduce latency variation as a result of the CRT beam fly‐time.

Visual stimuli comprised drifting (5 Hz) sine‐wave achromatic and cone‐isolating gratings of variable spatial frequency and contrast (stimulus duration 2 s; inter‐stimulus interval 2.5 s; 20–60 cycles presented) and temporal square‐wave intensity and/or chromaticity modulation of a spatially uniform circular field (pulse, duration 200 ms; 100 repetitions; inter‐stimulus interval 600 ms). The field size for grating and pulsed stimuli was 1–8° in diameter. Short‐wave sensitive (S) and medium/long‐wave sensitive (ML) cone‐isolating pulses and gratings were constructed by convolving marmoset cone spectral sensitivity with the spectral distribution of the monitor phosphors (Tailby et al. 2008; Pietersen et al. 2014). During the measurement of maintained activity, the screen was held at the mean luminance. Between eight and 20 epochs of data were recorded, each of 40 s duration, with a 20 s inter‐epoch interval.

Cell classification and histological reconstruction

The P, M and K cells were distinguished by their responses to the stimuli described above. The P cells show a sustained response to maintained contrast and linear contrast–response function. The M cells show a transient response to maintained contrast, saturating contrast response function and response phase advance from intermediate to high contrast levels (Wiesel & Hubel, 1966; Dreher et al. 1976; Kaplan & Shapley, 1986; Kaplan & Benardete, 2001). The K cells showed a variety of response properties, including colour‐selective ‘blue‐on’ (K‐bon), ‘blue‐off’ (K‐boff), ‘orientation‐selective’ (K‐o) cells and ‘suppressed‐by‐contrast’ (K‐sbc) receptive fields (Martin et al. 1997; Szmajda et al. 2006; Tailby et al. 2007; Roy et al. 2009; Cheong et al. 2013). The majority of K cells reported in the present study (43/60, 71%) are K‐bon cells.

At the conclusion of the experiment, an i.v. overdose of 120 mg kg⁻¹ pentobarbital sodium (Lethabarb; Vibac, Milperra, NSW, Australia) was given and then the animal was transcardially perfused with 0.9% saline followed by 4% paraformaldehyde. Recording track location was reconstructed as described in detail in previous studies (White et al. 2001; Szmajda et al. 2008). Anatomical location was confirmed for 55% (70/128) of cells. In cases where track location was not determined, the receptive field properties, eye dominance, encounter position and response characteristics described above were used as criteria.

Based on the combined anatomical and physiological criteria described above, one K cell was in the (ventral‐most) layer K1, two cells were in K2 (between the M layers), 11 cells were in K3 (between the M and P layers) and 13 were in K2 (between the internal and external P layers). Two K cells were located ‘ectopically’ in P layers and two in M layers . The laminar location of 15 K cells could not be determined unequivocally.

Colour vision phenotype

The red–green colour vision phenotype was estimated during the recording session from the pattern of responses to red–green contrast varying stimuli as described in detail previously (Blessing et al. 2004; Forte et al. 2006). Four animals (three male, one female) showed a 543 nm dichromatic phenotype; four animals (three male, one female) showed a 556 nm dichromatic phenotype; one male animal showed a 563 nm dichromatic phenotype; and one female animal showed a 556/563 nm trichromatic phenotype. The phenotype of the remaining three animals (one male, two female) could not be determined with confidence because insufficient tests were run. No clear differences in the results reported hereinafter were seen on comparing male and female animals, nor on comparing dichromatic animals with the single trichromatic animal.

Data analysis

Offline analysis was performed using Matlab (MathWorks, Natick, MA, USA). Spectral power in V1 LFP was estimated by multitaper analysis (taper parameter 10; moving window width 3 s, step size 0.3 s) using the Chronux toolbox for Matlab (Bokil et al. 2010). Power was normalized to the mean power between 80 and 100 Hz, and average power for each frequency band (delta: 1–4 Hz, theta: 5–8 Hz, alpha: 8–12 Hz and beta: 12–30 Hz) was calculated. The filtered signal was classified as showing high delta power (referred to as the synchronous cortical state) if the RMS delta power was more than 2 SD above the mean value for >3 s, and low delta power (referred to as the asynchronous cortical state) otherwise. Because we are recording from anaesthetized animals, the EEG frequency spectrum is biased to subgamma frequencies (Steriade, 2003; Buzsáki & Draguhn, 2004). Thus, our definition of the asynchronous cortical state includes more low‐frequency power than would be expected in studies of waking brain activity.

Maintained spike rates for P, M and K cells were calculated using the same moving window parameters as used for the LFP analysis. Spike phase was analysed by taking the instantaneous phase from the Hilbert transform of filtered delta LFP at the time of spike occurrence (Le Van Quyen et al. 2001). Rayleigh's test for non‐uniformity of circular data was used to estimate the strength of phase locking (Berens, 2009). The timing of LGN cell spike rate and V1 delta frequency power was subjected to Granger causality analysis using the bsmart toolbox for Matlab (Cui et al. 2008). A transition from low to high LGN spike rate was defined as a spike rate increase of at least 20 imp s⁻¹ for at least 2 s. Spike rate and delta frequency power were first Z‐scored, and then Granger causality was measured using model order of 3 (derived by Akaike information criterion).

The relationship between K cell spike rate and responses to high contrast stimuli were measured using pulse stimuli and receiver operator characteristic (ROC) analysis. Performance curves for ROC analysis were generated by comparing mean spike rate in 350 ms before and 200 ms following stimulus onset. The area under the curve (AUC) of the generated curve was taken as a measure of performance (Bradley, 1997).

Statistical analysis

Data are presented as the mean ± SD unless stated otherwise. Multiple group comparisons were made using Kruskal–Wallis non‐parametric analysis of variance with post hoc K‐Bonferroni‐corrected multiple pairwise comparison (Matlab functions kruskalwallis and multcompare); subsequently referred to as Kruskal–Wallis and Multcompare tests. Two‐group comparisons were made using the Wilcoxon non‐parametric rank sum test; subsequently referred to as the Wilcoxon test. All significance values are reported to two decimal places.

Results

Spike rate characteristics of P, M and K cells

The responses of 128 LGN cells (43 P, 25 M, 60 K) are reported. Not all cells were used in every analysis. Figure 2 shows examples of visually evoked and maintained discharge properties of P, M and K cells. The most readily characterized K cell type in macaque and marmoset monkeys is the blue‐on/yellow off (K‐bon) cell class (White et al. 1998; Szmajda et al. 2006; Roy et al. 2009); accordingly, the majority of K cells included in the present study (43/60, 71%) are K‐bon cells. Figure 2 A shows, for a K‐bon cell, the responses to pulse stimuli (left) and a spatial frequency tuning curve for S‐cone isolating gratings. This cell shows response properties typical of K‐bon cells, including a vigorous response to S‐cone contrast increments, complete response suppression to ML‐cone contrast increments and mild bandpass spatial tuning for S‐cone gratings (White et al. 1998; Tailby et al. 2008). Figure 2 B shows responses of an M cell in the same format as Fig. 2 A; this cell shows vigorous responses to achromatic and ML‐cone increments, is weakly suppressed by S‐cone increments, and demonstrates a saturating contrast–response function (Fig. 2 B right, inset) for achromatic gratings. The P‐off cell (Fig. 2 C) shows maintained responses to achromatic contrast decrements, a negligible response to S‐cone decrements, and linear contrast sensitivity for achromatic gratings.

Figure 2. Cell classification and brain state measurement.

A, response profile of the most commonly recorded K cell type: the cone opponent ‘Blue‐ON’ (K‐bon) cell. Visual field eccentricity 5.5°. Left, peristimulus time histograms (PSTH) of responses to 200 ms square wave pulse stimuli: achromatic increment (ACH+), short‐wave cone‐isolating increment (S+) and medium/long‐wave cone‐isolating increment (ML+). Stimulus duration is shown beneath each PSTH. Right, spatial frequency tuning for S‐cone‐isolating drifting gratings. B, responses of a magnocellular on‐centre cell (M‐on) in same format as in (A). Visual field eccentricity 6.7°. Tuning curve shows responses to achromatic gratings. Inset: contrast–response function for optimum spatial frequency gratings. C, response of a parvocellular off‐centre cell (P‐off) in same format as in (B). Visual field eccentricity 6.9°. PSTHs show a response to decrement pulses. D, simultaneous recording of the LFP in V1 and spike rate of the K cell shown in (A). The black trace (upper) shows the unfiltered LFP; the blue trace (centre) is the LFP bandpass filtered for delta frequencies (1–4 Hz). Tick marks on the lower panel show individual K cell action potentials; grey bars show the PSTH. Note the increase in spike rate during periods of cortical asynchrony (low delta power). E, responses of the M cell in the same format as in (D). F, responses of the P cell in same format as in (D). [Color figure can be viewed at wileyonlinelibrary.com]

Figure 2 D and F shows recordings where these example cells were presented with a uniform grey screen (∼50 cd m⁻²). The upper traces show local field potential, bandpass filtered for delta frequencies (1–4 Hz). The vertical tick marks represent individual action potentials; the lower histograms show PSTHs of spike rates in 0.5 s bins. Around 15 s into the K cell recording, the spike rate increases markedly for ∼10 s (start marked with arrow). This increase is associated with reduced cortical synchrony, as indicated by the reduction of delta‐band amplitude in V1. Comparable changes in cortical synchrony have little or no effect on ongoing activity of the M cell (Fig. 2 E) or the P cell (Fig. 2 F).

Figure 3 shows that, as a population, K cells show comparable mean rate (Fig. 3 A) but higher spike rate variability (Fig. 3 B) than M cells (Fig. 3 C) and P cells (Fig. 3 D). For K cells, the mean variability ± SD is 5.1 ± 4.0; for M cells, the mean variability ± SD is 3.1 ± 1.7; for P cells, the mean variability ± SD is 2.7 ± 1.3, Kruskal–Wallis P < 0.01). For example, almost half (19/45; 42%) of K cells show an SD >5 imp s⁻¹, whereas only 3/43 (7%) of P cells and four of 25 (16%) of M cells do. Consistently, a K‐means cluster analysis (Matlab statistics toolbox) divides the K cell recordings into two groups with low (<5 imp s⁻¹) and high (>5 imp s⁻¹) SD (data not shown). This division could indicate the existence of discrete K cell classes with high or low intrinsic variability. More probably, however (see below), changes of V1 state occurred during some K‐cell recordings but not during others. Consistent with a previous study (Cheong et al. 2011), across the P, M and K cells sampled, we found no significant differences in average maintained spike rate (Fig. 2 A) between the three classes (K, 11.4 ± 7.1, n = 45; M, 11.6 ± 6.1, n = 25; P, 9.0 ± 4.6, n = 43; Kruskal–Wallis P = 0.15). Below, we examine the relationship between LGN cell spike rate and delta‐band power in V1 in more detail.

Figure 3. Comparison of spike rate variability.

A, average maintained spike rates across K, M and P cell populations. Each point shows average for one cell across 480–720 s of recording. Grey bars show the mean ± SD for each population. Average spike rate is not significantly different across the populations. B, histogram of variability in spike rate of K cells. Approximately half of the K cells show a SD >5 imp s⁻¹. C, four M cells show a SD >5 imp s⁻¹. D, three P cells show a SD >5 imp s⁻¹. [Color figure can be viewed at wileyonlinelibrary.com]

K cell spike rate fluctuations correlate with cortical synchrony

Figure 4 A shows an example 40 s recording, with spectral analysis of V1 LFP (Fig. 4 A, upper), K cell spike rate (Fig. 4 A, centre) and the average power ratio (Fig. 4 A, lower) over the delta (1–4 Hz, blue), theta (5–8 Hz, green), alpha (8–12 Hz, red) and beta (12–30 Hz, cyan) frequency bands. Power in each frequency band was normalized to the average power between 80 and 100 Hz. The lack of power in V1 LFP (lack of warm colours) above 10 Hz is consistent with deep surgical anaesthesia. Just before the 15 s mark, the K cell spike rate increases markedly. This increase is accompanied by reduced cortical synchrony, as indicated by (i) a loss of power in the lower frequency bands (lack of warm colours in the heat map) and (ii) a reduction in the LFP ratios, most notably in the delta band. We restricted our analysis to the relationship of delta band power with LGN spike rates, although inspection of Fig. 4 A shows that weaker relationships of spike rate to other LFP frequency bands may also be present.

Figure 4. Relation of spike rate to brain state.

A, moving window (3 s, 0.3 s steps) spectrogram of the LFP in V1 (upper), K cell spike rate (centre) and average power in LFP frequency bands (bottom) of a 40 s recording epoch (this cell is also shown in Fig. 2 A and D). In the spectrogram, warm colours indicate greater power. The LFP was normalized to the mean power at 80–100 Hz. Note the loss of warm colours in the lower frequencies between 10 and 15 s, shortly before an increase in K cell spike rate. B–E, four examples of the relationship between LGN cell spike rate (x‐axis) and relative delta‐band LFP power (y‐axis). All cells have a spike rate variability >5 imp s⁻¹ (Fig. 1 D). Each circle represents one windowed sample calculated as in (A). Both K cells (B and C) show a negative correlation between spike rate and LFP ratio, whereas the M and P cells (D and E, respectively) show no clear relationship. F–G, correlation coefficients derived from linear regression of LGN cell spike rate and delta LFP ratio. Grey bars show cells with a significant correlation; open bars show non‐significant correlations. Cells with high firing rate variability (SD >5 imp s^–1) (Fig. 3) are distinguished with dot markers in the histogram. Note that K cells with high variability cluster to the left of the histogram, indicating high negative correlation. [Color figure can be viewed at wileyonlinelibrary.com]

For each cell, between eight and 20 epochs (each of 40 s duration; see Methods) were recorded. Figure 4 B–E shows scatter plots for four example LGN cells with spike rate on the x‐axis and the simultaneously recorded delta frequency LFP ratio on the y‐axis. All of these cells showed high overall variability in spike rate (SD > 5). The K cells show a negative overall correlation, where the LGN cell spike rate is high when V1 is in an asynchronous state. The K cell in Fig. 4 B (K‐bon cell) shows a systematic decline in LFP ratio with increasing spike rates. The K cell in Fig. 4 C [K blue‐off (K‐bof) cell] by contrast shows two distinct modes in spike rate. Of 19 K cells that that displayed high variability in spike rate (SD >5), 16 showed a significant negative correlation with delta power (P < 0.01, Pearson correlation). The example M (Fig. 4 D) and P (Fig. 4 E) cells show no clear relationship with cortical state, despite the fact that these two cells showed, respectively, the greatest variability of all M and P cells in our dataset. The distribution of correlation coefficients derived from linear regression of LGN cell spike rate and delta LFP ratio is shown in Fig. 4 F (P‐ and M‐cells) and Fig. 4 G (K‐cells). Here, grey bars show cells with a significant correlation; open bars show non‐significant correlations. Cells with high firing rate variability (SD > 5) (Fig. 3) are distinguished with dot markers in the histogram. It is evident by inspection that the K cells with high variability are clustered at the left of the histogram and have high and significant negative correlation with delta power.

To further characterize the relationship between cortical state and LGN spiking activity, we next used Hilbert analysis to estimate delta band power (see Methods) and segmented the recording epochs into asynchronous (low‐delta) and synchronous (high‐delta) states. Across the dataset, almost equal periods of synchronous and asynchronous state were sampled (for low‐delta: K, 215 ± 15 s, n = 45; P, 200 ± 16 s, n = 43; M, 153 ± 11 s, n = 25; Kruskal–Wallis P = 0.02; for high‐delta: K, 163 ± 13 s; P, 163 ± 10 s; M, 173 ± 16 s; Kruskal–Wallis P = 0.80). The analysis yields two main results. First, as expected from the data so far, there is a selective link between K cell spike rate fluctuations and cortical state. Figure 5 A shows that cells with high spike rate variability (SD > 5) were predominantly recorded during asynchronous (low delta) cortical state, and Fig. 5 B shows that mean spike rate for almost all of these cells is greater in an asynchronous than in a synchronous cortical state. Of 19 K cells shown on these plots, 16 (84%) were recorded predominantly during asynchronous cortical state and 15 (79%) of these showed a higher maintained rate in an asynchronous state (chi‐squared P < 0.01 for both comparisons). Second, and perhaps more surprisingly, the behaviour of K cells is not completely uniform: similar to P and M cells, some K cells with overall low variability (SD ≤5) were predominantly recorded during asynchronous cortical state (Fig. 5 C) but showed no clear link between spike rate and cortical state (two examples are marked with arrows in Fig. 5 C and D). These ‘unlinked’ K‐cells were not obviously restricted to one region of LGN recordings or one visual response type. Nevertheless, we can conclude that epochs of high spike rate in K cells are predominantly linked to an asynchronous cortical state, whereas spike rates of almost all P and M cells are not influenced by cortical state. The distribution of state times for high and low delta states is shown in Fig. 5 E and F. Both distributions are heavily skewed to durations below 10 s [low delta mean 5.2 ± 4.4 (n = 3292); high delta mean 6.2 ± 4.1 (n = 1459)].

Figure 5. Comparison of cells showing high and low variance in maintained activity.

A, scatterplot, for high variance cells, of the cumulated recording time where cortex was in a synchronous (high delta) or asynchronous (low delta) state. B, comparison, for high variance cells, of the mean discharge rate in a high delta and low delta state. Note that almost all K cells show an increased spike rate in an asynchronous cortical state. C, comparison of cells showing low variance, in same format as in (A). Arrows indicate examples of K cells recorded predominantly in an asynchronous cortical state. D, comparison of spike rates, in same format as in (B). Note that spike rates of the marked K cells show negligible change. E, histogram of high delta state durations. F, histogram of low delta state durations. [Color figure can be viewed at wileyonlinelibrary.com]

Cortical asynchrony precedes K cell spike rate increases

An obvious question arising is whether V1 state changes occur before or after K cell spike rate changes. Accordingly, we employed directed Granger causality analysis (Cui et al. 2008). The LFP and spike rates signals were measured in the same way as for the data shown in Fig. 4, and then both signals were Z‐transformed across the 40 s recording epoch. Figure 6 A shows two examples of low‐to‐high K cell spike rate transition (thin line) accompanied by high‐to‐low delta LFP power transition (Fig. 6 A dotted line). Time is shown relative to the start of the recording epoch. To enable the Granger analysis, we needed to identify transitions that were long‐lived but did not fall off the borders of the recording epoch. In total, 27 clearly defined examples of transitions from low‐to‐high spike rate (defined as explained in the Methods section) were identified in 7 K cells. For each transition period, the mean causality in the 0–0.5 Hz range was calculated for LGN→V1 and V1→LGN directions. The result is shown as a scatterplot in Fig. 6 B. Each recorded cell is indicated by a different symbol. With one exception, the points lie above or close to the unity line, indicating that the asynchronous cortical state is a better predictor of LGN spike rate than vice versa. Accordingly, the mean V1→LGN causality (1.4 ± 0.12, n = 27) was higher than LGN→V1 causality (0.9 ± 0.09, Wilcoxon P < 0.01). Analysis of high‐to‐low spike rate transitions yielded a consistent result (mean V1→LGN causality 1.4 ± 0.6, n = 15; LGN→V1 causality 0.9 ± 0.5, Wilcoxon P < 0.01). These results are broadly consistent with propagation of brain state V1 to LGN, although it is important to note that the V1 to LGN feedback pathway is part of a complex loop system comprising multiple pathways within and between the LGN, V1 and thalamic reticular nucleus (McCormick, 1989; Contreras & Steriade, 1995; Sherman & Guillery, 1996; Callaway, 2005; Crunelli & Hughes, 2010). We will return to the question of causality later in the Results; for now, we reiterate the main result that changes in cortical state are a better predictor of changes in K cell state than the reverse. For expediency, we did not attempt more complex analyses of causality (Hu et al. 2016).

Figure 6. Granger causality analysis.

A, two examples of a cortical state change. The thick grey line shows the maintained spike rate of a single K cell; the broken line shows the LFP delta‐band power ratio. Data are Z‐scored to allow direct comparison. Time on the x‐axis is the relative time from the start of the 40 s recording epoch. B, Granger causality of 27 transitions recorded in 7 K cells. Each symbol indicates a different K cell. The black line is the unity line. The x‐axis is mean directed causality LGN to V1 over the transition period; the y‐axis is the mean directed V1 to LGN. On average, most transitions fall above the unity line indicating the V1 state predicts K cell spike rates.

K cell spike rate fluctuations are independent of spike bursts in LGN

Spike bursts in dorsal thalamic nuclei typically comprise a brief (50–100 ms) pause in spike activity followed by a short train of high‐frequency spikes. Spike bursts in thalamus are more common under synchronous (e.g. sleeping) than asynchronous (e.g. waking) cortical states (Contreras & Steriade, 1995; Ramcharan et al. 2000b; Sherman & Guillery, 2006). We therefore expected that the frequency of spike bursts in marmoset LGN would be greater in a synchronous (high‐delta) state than in an asynchronous (low‐delta) state. Surprisingly, however, we found this not to be the case. We defined bursts using criteria very close to the liberal criteria used by Ramcharan et al. (2000): a minimum of 50 ms pause before burst onset and less than 6 ms between burst spikes. Bursts of two spikes are included in the criterion. Plotting the interval before spike onset against the interval after spike occurrence for a single cell during synchronous (Fig. 7 A) and asynchronous (Fig. 7 B) state shows a very similar pattern. Spikes that are part of a burst are shown as square symbols; all other recorded spikes are shown as black points. In this example, during periods of cortical synchrony (Fig. 7 A), 101 bursts were detected, whereas, during asynchronous periods (Fig. 7 B), 80 bursts were detected. Hilbert phase analysis revealed no selective effect of cortical state on delta phase for burst spikes (Fig. 7 C and D).

Figure 7. Analysis of spike bursts.

A, scatterplot of interspike interval (ISI) during synchronous (high delta) cortical state. Square symbols show spikes that were classified as part of a burst. Black dots show non‐burst spikes. Solid grey lines show thresh‐ old ISIs of spikes within the burst (6 ms). Broken grey lines show threshold before‐spike interval (50 ms). B, ISI scatterplot of the same cell during an asynchronous (low delta) state. No clear difference to a synchronous state is evident. C–D, delta phase of the burst during synchronous (high delta, C) and asynchronous (low delta, D) cortical periods. There is no preferred phase where bursts occur. E, number of bursts per second for K, P and M cells. Significantly fewer bursts were recorded in K cells (P < 0.02). F, no difference was recorded between the instance of bursts during synchronous (high delta) and asynchronous (low delta) cortical periods for K, P or M cells. [Color figure can be viewed at wileyonlinelibrary.com]

Overall, K cells show the lowest instance of bursts (0.29 ± 0.4 burst s⁻¹, n = 45) compared to P (0.63 ± 0.6 burst s⁻¹, n = 43) and M cells (0.43 ± 0.3 burst s⁻¹, n = 25, Kruskal–Wallis P < 0.01) (Fig. 7 E). Comparison of burst rate revealed no difference on comparing synchronous and asynchronous cortical state (Fig. 7 F) for any cell population (K cells: 0.29 ± 0.4 vs. 0.27 ± 0.3 bursts s⁻¹, n = 45, paired Wilcoxon P = 0.95; P cells: 0.63 ± 0.6 vs. 0.62 ± 0.5 bursts s⁻¹, n = 43, P = 0.5; M cells: 0.45 ± 0.3 vs. 0.43 ± 0.3 bursts s⁻¹, n = 25, P = 0.88). There was no significant difference in cortical phase of the bursts between cell types (K: −2.0 ± 1.3 rad; M: 2.5 ± 1.3 rad; P: 2.6 ± 1.3 rad, circular Kruskal–Wallis P = 0.91). As might be expected from the almost identical appearance of the scatterplots in Fig. 7 A andB, other measures that we tested (e.g. burst fraction) and repeat analysis using conservative (100 ms pause, 4 ms between spikes) and very liberal (0 ms pause, 6 ms between spikes) burst criteria, as well as comparison of K cells showing high (SD > 5) and low (SD < = 5) spike rate variability, similarly showed negligible differences between a synchronous and asynchronous cortical state (data not shown).

These results show that bursts are less frequent in the maintained activity of K cells than in the maintained activity of M or P cells, and that there is no clear relationship between cortical delta state and LGN cell burst frequency or timing. This conclusion is limited by the fact that (as noted above in the Methods), our definition of asynchronous state includes more EEG power at subgamma frequencies than would be present in studies of waking animals.

K cell spike rate fluctuations interact additively with sensory stimuli

All of the results described so far are from the analysis of ongoing (maintained or ‘spontaneous’) activity recorded in the absence of patterned visual stimuli. These analyses do not reveal how K cell spike rate fluctuations could interact with, and therefore modulate, sensory signals. Transmission of retinal spikes through LGN is dependent on visual stimulus parameters such as contrast and temporal frequency (Kaplan et al. 1987; Sincich et al. 2007), as well as cortical state (Kaplan et al. 1993; Li et al. 1999; Funke & Eysel, 2000). Specifically, we asked whether the interaction between sensory signal and K cell spike rate fluctuations is additive or multiplicative. Figure 8 A shows, schematically, the predicted effect of additive and multiplicative noise interaction with visually evoked signals. Expressed as a function of noise amplitude, additive noise produces a uniform response amplitude increase (change in elevation; left scatterplot), whereas multiplicative noise produces an amplitude‐dependent increase (change in gain, right scatterplot). Figure 8 B shows, for a K‐bon cell, the maintained response rate before stimulus (Fig. 8 B, grey square markers) and the evoked response rate during stimulus (Fig. 8 B, blue triangle markers) across 100 repetitions of a 200 ms S‐cone selective pulse. Over the course of ∼8 min of data collection, the maintained spike rate for this cell varied between 0 and 80 imp s⁻¹. It is immediately obvious that the evoked responses run above and almost parallel to the maintained responses. Figure 8 C shows peristimulus time histograms of evoked responses collected during all trials (Fig. 8 C, left) or trials where the maintained spike rate before stimulus onset was high (Fig. 8 C, centre) or low (Fig. 8 C, right). Figure 8 D shows a scatter plot of the maintained rate before stimulus onset (x‐axis) against the evoked spike rate during stimulus presentation (y‐axis) across 126 stimulus presentations. The thick black line shows mean regression (y = 1.1x + 13). The slope of the fit is close to the slope of the unity line (Fig. 8 D, thin grey line); thus, the evoked visual response in this K cell is superimposed on a variable baseline (Fig. 4 B and G). Comparable results were obtained for almost all K cells tested. Figure 9 A shows linear fits for individual K cells (Fig. 9 A, thin grey lines), as well as population mean regression (y = 0.9x + 32 imp s⁻¹, n = 20) (Fig. 9 A, thick magenta line). Thus, across our sample, K cells increase their spike rate on average by 32 imp s⁻¹ across the 200 ms preferred stimulus presentation, and this increase is independent of spike rate before stimulus onset, indicating additive signal and noise combination.

Figure 8. Signal‐to‐noise analysis.

A, predicted effects of signal interaction under additive or multiplicative noise. Additive noise (left) predicts a uniform increase in signal + noise response amplitude (blue points on scatterplot). Multiplicative noise (right) predicts an amplitude‐dependent increase in signal × noise. B, evoked response amplitude of a K‐bon cell before (grey) and during (blue) repeated presentations to a 200 ms S‐cone pulse. C, peristimulus time histograms of all stimuli (left) and during periods of high (centre) and low (right) maintained activity. A dark bar indicates the stimulus duration, which was also the spike rate measurement window. D, same data as in (B) but expressed as the mean spike rate in 350 ms before stimulus onset (x‐axis) against the mean spike rate during the 200 ms stimulus. Stimulus is a S‐cone isolating increment (S‐ON). Open and filled points, respectively, show stimulus presentations where the spike rate before stimulus onset was below or above 20 imp s^–1. The black line shows the linear regression through all points; formula on the bottom right. The grey line is the unity line. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 9. ROC analysis.

A, linear regression of K cell responses to preferred stimulus polarity, calculated as shown in Fig. 8 D. Grey lines show individual K cells; the magenta thick line shows the mean regression line; the back line shows the unity line. The formula at the bottom right is the formula from the mean regression. B, performance curves from ROC analysis from the same data as in Fig. 8A. The light broken line is the performance from points with a high spike rate before stimulus onset; the thin solid line is the performance of points with a low spike rate before stimulus onset. The number in the legend is the AUC expressed as a percentage. C, ROC analysis for all K blue on cells (n = 20) to S‐cone increments. Triangles and squares represent individual cells (square = ‘blue‐off’, triangle = ‘blue‐on’). Grey squares show the mean ± SEM. D, response of the same cell as in Fig. 8, although for S‐cone isolating decrement (S‐off) stimulus. E, performance curves from ROC analysis for responses to anti‐preferred stimuli for the same cells as in A–C. F, ROC analysis for anti‐preferred stimulus. [Color figure can be viewed at wileyonlinelibrary.com]

We next applied ROC analysis to cell responses. Technical limitations prevented us from storing V1 LFP signals during these recordings. We therefore used the spike rate before stimulus onset as a proxy for cortical state. This approach is justified because we are investigating the relationship of maintained spike rate with stimulus detectability. The analysis does not depend on cortical state per se, nor does it make any inference about cortical state. Each stimulus trial was classified as showing low or high spike rate before stimulus onset. A criterion at 20 imp s⁻¹ was chosen because 95% of pre‐stimulus spike rate samples fell below this value in cells that showed low spike rate variability (SD <5; 2356 trials recorded from 23 K cells) (Fig. 5 B and F). The ROC was then calculated for these two spike rate states. Figure 9 B shows ROC analysis of the cell shown in Fig. 8. We used the AUC as a measure of stimulus detectability. In this example, there is negligible difference in detectability if the spike rate before stimulus onset was low (AUC 86%) or high (AUC 83%). The same conclusion can be drawn from the other K cells analysed (Fig. 9 C) (n = 20) in that there is negligible difference in detectability between low or high spike rate before stimulus onset (0.94 ± 0.08 vs. 0.91 ± 0.11, Wilcoxon P = 0.10).

When a cell is presented with an anti‐preferred stimulus, its spike rate (by definition) decreases. Figure 9 D and E shows the ROC analysis for an anti‐preferred stimulus (S‐cone contrast decrement) of the cell shown in Fig. 8. It is apparent that the signal–noise interaction is better expressed as a change of slope (multiplicative signal–noise interaction). Consistently, the mean regression slope for anti‐preferred stimuli was lower than for preferred stimuli (0.2 ± 0.2 vs. 0.9 ± 0.3, paired Wilcoxon P = 0.00, n = 20). Under high maintained rates, the dynamic response range is thus improved (i.e. less rectified) relative to low maintained rates. Consequently, high maintained rates can improve the neurometric detectability of anti‐preferred stimuli (AUC 84% for low spike rate; AUC 99% for high spike rates) (Fig. 9 E). The same pattern emerges from the population data (Fig. 9 F). Trials with a high maintained rate before onset are better detectors of evoked spike rate decrements (0.97 ± 0.06) than trials with a low spike rate (0.84 ± 0.12, Wilcoxon P = 0.00, n = 20). We conclude that high spike rates in K cells do not, in principle, degrade the capacity to detect stimuli, and may even improve that capacity.

Delta‐frequency phase‐locking in M, P, and K populations shows cortical state dependence

The complementary changes in cortical state and K cell spike rate described above take place across multisecond timescales. We previously showed that synchronous activity in pairs and ensembles of K cells can also be detected at subsecond timescales (Cheong et al. 2011). Therefore, we investigated whether LFP delta band activity is linked to spike rates at the subsecond timescale of individual delta‐band waves (1–4 Hz). Figure 10 A shows a raw V1 local field potential trace (Fig. 10 A, black, upper), delta band filtered trace (Fig. 10 A, blue, centre) and simultaneously recorded spikes from a single K cell (Fig. 10 A, blue triangles, lower). We used Hilbert transform to calculate the instantaneous phase of the delta band signal for each spike across the 40 s recording epochs for K cells (n = 45), M cells (n = 25) and P cells (n = 43). Cortical state was classified as asynchronous (low delta power) or synchronous (high delta power) as described in the Methods. Figure 10 B shows an example phase distribution for spikes of a single K‐bon cell during a total 176 s recording made during high‐delta state. We used Rayleigh's non‐uniformity test to estimate the strength of phase locking of spikes to delta waves. In this example, the spike probability is highest slightly before the delta wave peak (mean angle ± variance, –1.44 ± 0.9 rad, vector length = 0.1, Rayleigh's test for non‐uniformity P < 0.01).

Figure 10. Delta‐frequency phase locking.

A, example of delta‐frequency spike timing analysis. Black trace (upper) is the raw LFP; the blue trace (centre) is the same data bandpass filtered for delta frequencies (1–4 Hz). Blue triangles represent action potentials from a single K‐blue on cell. B, example distribution of non‐uniform phase relationship between LGN cell spike rate and V1 delta waves. The histogram shows the instantaneous phase taken from a Hilbert transform of the delta‐filtered LFP at the time of each spike. The solid line is an illustration of the voltage–phase relationship for a single delta wave. C, fraction of phase‐locked cells during low and high delta periods. D, polar plots of all cells that showed significant phase‐locking during high delta periods. The thick black line shows the mean vector direction. [Color figure can be viewed at wileyonlinelibrary.com]

Across all LGN cell types, we found that ∼50% of each population (K: 22/45, M: 14/25, P: 22/43) shows significant phase locking during the high‐delta state (Fig. 10 C). Figure 10 D is a vector plot of preferred phase angle and bias strength for the ∼50% of cells showing phase locking in high‐delta state (thin lines) and mean (thick line) vector (mean angle ± variance, K cells: −2.36 ± 0.6 rad, vector length = 0.4; M cells: 2.72 ± 0.39 rad, vector length = 0.61; P cells: −2.97 ± 0.55 rad, vector length = 0.45). Spiking in most LGN cells is phase biased around the trough of V1 delta waves. During an asynchronous cortical state, for K and P cells, the angle of phase locking remained similar, although the mean angle for M cells moved closer to the peak of V1 delta waves (mean angle ± variance, K cells: −2.09 ± 0.43 rad, vector length = 0.57; M cells: −0.36 ± 0.87 rad, vector length = 0.13; P cells: 2.84 ± 0.44 rad, vector length = 0.56).

The proportion of phase locked cells decreases to ∼25% for all LGN populations under an asynchronous cortical state (K cells: 11/45, M cells: 8/25, P cells: 8/43) (Fig. 10 C). This last result shows that, although only K cell spike rates are linked to slow changes in a cortical state, delta‐band thalamocortical synchrony is lower in all LGN divisions under an asynchronous cortical state than under a synchronous cortical state.

Discussion

We found that, in a synchronous cortical state, activity throughout all (P, M and K) divisions of the LGN is entrained to delta‐band cortical oscillations. Switches from a synchronous to asynchronous state are accompanied by spike rate changes in K pathway cells, as well as by decreased delta entrainment in all LGN divisions. The switches from a synchronous to asynchronous cortical state precede (normally by 1 – 3 s) the switches from a low to high K cell spike rate, implying that an asynchronous cortical state could be the cause of spike rate changes in the K layers. Furthermore, switches in K cell spike rates do necessarily not come at cost of faithful visually‐evoked signal transmission.

We found uniform phase‐locking to delta waves across the LGN under a synchronous cortical state. Entrainment of thalamic and cortical activity is consistently observed across species and dorsal thalamic nuclei studied so far in non‐primate mammals (Contreras & Steriade, 1995; Rigas & Castro‐Alamancos, 2007; Crunelli & Hughes, 2010) and we show (for the first time to our knowledge) a consistent pattern of delta‐band entrainment in primate LGN. We concentrated on delta‐band cortical activity, although LGN coherence to other temporal frequency bands would probably show a similar pattern (Lőrincz et al. 2009). Thalamocortical coherence at low frequencies has been specifically linked to low threshold calcium channels and thalamic burst activity (Steriade et al. 1993; von Krosigk et al. 1993; Ramcharan et al. 2000a; Llinás & Steriade, 2006; Crunelli & Hughes, 2010). In our experiments, however, we did not see a clear relationship between LGN burst activity and the phase of delta‐frequency oscillations (Fig. 7) and, overall, we found little burst activity in marmoset LGN.

Ramcharan et al. (2000a) found an overall low frequency of bursts in the waking macaque LGN compared to the somatosensory thalamus. It was suggested that relatively high levels of synaptic bombardment from retinal inputs could serve to keep LGN cells in a more depolarized state than relay cells in other dorsal thalamic nuclei. On the other hand, Livingstone & Hubel (1981) showed increased burst activity associated with delta‐band EEG activity in the LGN of cats in transition from a waking to sleeping state (see also Kaplan et al. 1987; Funke & Eysel, 2000). Thus, in addition to differences between thalamic nuclei, there may be differences between carnivores and primates in the relation of LGN bursting and cortical state.

We found selective covariance of K cell activity with cortical state, where asynchronous cortical states are linked to high spike rates in K cells. These results are obtained in a much larger sample of LGN cells (60 K cells, 43 P cells, 25 M cells) than was the case for the results reported by Cheong et al. (2011) who measured 27 LGN cells in total. In addition, our Granger causality analysis (Fig. 6 A and B) suggests that slow changes in brain state propagate from V1 to K cells in LGN. If this link is causal, then the asynchronous cortical state is promoting increased spike rates in part of the afferent thalamic input streams. On the other hand, Tan et al. (2014) showed that increased spike rates in thalamic inputs can cause the cortex to shift from a synchronous to asynchronous state. We speculate that two processes act as a positive feedback circuit, where loss of synchrony in the cortex raises the K cell spike rate, which in turn leads to further cortical desynchronization. In this way, the K pathway could be viewed as part of a circuit controlling switches between synchronous and asynchronous cortical states. Of course, our observations are made under anaesthesia, and so the extent to which these physiological principles would apply to conscious visual processing is not yet known.

In addition to showing widespread terminations in supragranular layers of V1 (Blasdel & Lund, 1983; Fitzpatrick et al. 1983; Casagrande et al. 2007), K pathways are distinct to P and M pathways in providing inputs to non‐striate cortical visual areas (Bullier & Kennedy, 1983; Sincich et al. 2004; Warner et al. 2010). These anatomical features have led to the thalamocortical synchrony hypothesis (Jones, 2001, 2009) where cortical activity generated by K pathways could co‐ordinate activity within and between visual cortical area and generate coherent thalamocortical oscillations. Our cortical recordings were limited to the V1, and we do not have direct evidence that the K pathway is involved in controlling the cortical state. Our results are nevertheless consistent with the thalamocortical synchrony hypothesis because they show that slow cortical state changes are more tightly linked to K pathway activity than to P and M pathway activity.

Mechanisms

There is anatomical (Lund et al. 1975; Fitzpatrick et al. 1994; Usrey and Fitzpatrick, 1996) and physiological evidence (Briggs & Usrey, 2007, 2009) indicating a parallel organization of cortico‐geniculate pathways. Thus, there are anatomical bases for selective propagation of the cortical state to K cells but not P or M cells. Furthermore, there is anatomical and physiological evidence for the merger of K, P, and M afferent pathways at early stages of cortical processing (Yoshioka et al. 1994; Cottaris & DeValois, 1998; Vidyasagar et al. 2002; Callaway, 2005). In this way, activated K cells and intrinsic cortical circuitry (Fröhlich et al. 2006) could support asynchronous cortical activity. This scenario does not rule out a role of the connections of K layers to subcortical centres regulating eye movements, vigilance and attention (Harting et al. 1978; Casagrande, 1994; Bickford et al. 2000). For example, the basal nucleus of Meynert has widespread diffuse cholinergic connections with the cortex, as well as the thalamus, and could initiate or propagate changes in cortical state (Semba, 2000; Steriade, 2004).

Summary

The traditional view of the LGN as a relay structure has given way to the modern view of the LGN comprising one participant in a group of thalamocortico‐thalamic loops that serve perception and cognition (Jones, 2001; Sherman & Guillery, 2006; Steriade, 2006). In the present study, we show that low‐frequency changes in K cells reflect changes in thalamocortical coherence across the LGN. It is known that sensory signals at the LGN level can be influenced by attention processes (O'Connor et al. 2002; McAlonan et al. 2008; Saalmann & Kastner, 2011; Jiang et al. 2015). Thus, a change in brain state and loss of phase locking between LGN cells and cortex might indicate a first step in modulating sensory information in all LGN layers. It would obviously be of interest to know whether the result of the present study with respect to anaesthetized marmosets also apply to behavioural state transitions in awake animals.

Additional information

Competing interests

The authors declare that they have no competing interests.

Author contributions

ANJP, SC, PRM and SGS designed the research. ANJP SC, PRM and SGS performed the research. ANJP, SC, BM, PG, PRM and SGS analysed the data. ANJP and PRM drafted the paper. All authors revised the paper critically for important intellectual content. All authors approved the final version submitted for publication.

Funding

Funding for the present study was provided by Australian National Health and Medical Research Council Grants 1027913 and 1005427, as well as Australian Research Council grants CE140100007 and DP160104316.